Machine-Learning-Based Predictive Maintenance for Sustainable Manufacturing Process Monitoring Using Open Industrial Benchmark Data

Abstract

Introduction: Modern manufacturing systems increasingly depending on continuous process data to improve reliability, reduce downtime, and support efficient resource use. Though, many manufacturing environments continue to rely on reactive or schedule-based maintenance, which can result in unexpected equipment failures, unnecessary maintenance interventions, and inefficient use of operational resources. This ongoing study develops an explainable machine-learning framework for predictive maintenance using open industrial benchmark data.

Methods: The study will use the AI4I 2020 Predictive Maintenance Dataset to develop a supervised classification pipeline for predicting machine failure.  Logistic regression, random forest, gradient boosting, and extreme gradient boosting models will be compared using accuracy, precision, recall, F1-score, ROC-AUC, and PR-AUC.  Since machine failures represent a minority class, particular attention is given to failure recall, PR-AUC, missed failures, and false maintenance alerts.  SHAP analysis will be used to inspect the direction and scale of feature contributions, while permutation feature importance will assess the overall predictive importance of process variables.  The sustainability-related operational implications will be measured through measurable indicators, including missed-failure reduction, unnecessary maintenance alerts, maintenance efficiency, and potential downtime avoidance.

Results: The analysis is designed to conclude which model provides the most effective balance between identifying equipment failures and limiting unnecessary maintenance alerts.  The explainability analysis will identify the process conditions most strongly associated with predicted machine-failure risk.  The final empirical results will be reported in the completed conference study.

Conclusion: The study proposes a reproducible framework integrating predictive performance, explainability, and maintenance evaluation. Although AI4I 2020 cannot represent complex manufacturing environments, it provides a controlled benchmark. The study does not directly measure environmental sustainability; instead, it examines how failure detection and fewer unnecessary interventions may improve maintenance planning, reduce downtime, and conserve resources. Future research should validate the framework with real-world industrial datasets.